This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
NBER WORKING PAPER SERIES
CAN CHANGING ECONOMIC FACTORS EXPLAIN THE RISE IN OBESITY?
Charles J. CourtemancheJoshua C. Pinkston
Christopher J. RuhmGeorge Wehby
Working Paper 20892http://www.nber.org/papers/w20892
NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue
Cambridge, MA 02138January 2015
Monica Deza, Robert Kaestner, Rusty Tchernis, Nathan Tefft, and seminar participants at GeorgiaState University, University of Illinois-Chicago, the American Society of Health Economists BiennialConference, and the 5th Annual Meeting on the Economics of Risky Behaviors provided valuablefeedback. We thank Xilin Zhou and Antonios Koumpias for excellent research assistance. Ruhm thanksthe University of Virginia Bankard Fund for financial support for this research. The views expressedherein are those of the authors and do not necessarily reflect the views of the National Bureau of EconomicResearch.
NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies officialNBER publications.
Can Changing Economic Factors Explain the Rise in Obesity?Charles J. Courtemanche, Joshua C. Pinkston, Christopher J. Ruhm, and George WehbyNBER Working Paper No. 20892January 2015JEL No. I12
ABSTRACT
A growing literature examines the effects of economic variables on obesity, typically focusing on onlyone or a few factors at a time. We build a more comprehensive economic model of body weight, combiningthe 1990-2010 Behavioral Risk Factor Surveillance System with 27 state-level variables related togeneral economic conditions, labor supply, and the monetary or time costs of calorie intake, physicalactivity, and cigarette smoking. Controlling for demographic characteristics and state and year fixedeffects, changes in these economic variables collectively explain 37% of the rise in BMI, 43% of therise in obesity, and 59% of the rise in class II/III obesity. Quantile regressions also point to large effectsamong the heaviest individuals, with half the rise in the 90th percentile of BMI explained by economicfactors. Variables related to calorie intake – particularly restaurant and supercenter/warehouse clubdensities – are the primary drivers of the results.
Charles J. CourtemancheGeorgia State UniversityAndrew Young School of Policy StudiesDepartment of EconomicsP.O. Box 3992Atlanta, GA 30302-3992and [email protected]
Joshua C. PinkstonEconomics DepartmentCollege of BusinessUniversity of LouisvilleLouisville, KY [email protected]
Christopher J. RuhmFrank Batten School ofLeadership and Public PolicyUniversity of Virginia235 McCormick Rd.P.O. Box 400893Charlottesville, VA 22904-4893and [email protected]
George WehbyDepartment of Health Management and PolicyCollege of Public HealthUniversity of IowaN248 CPHB, 105 River StreetIowa City, IA 52242and [email protected]
1
I. Introduction
Obesity, defined as a body mass index (BMI) of at least 30, leads to adverse health
conditions such as heart disease, diabetes, high blood pressure, and stroke (Strum, 2002).1 The
adult obesity rate in the United States skyrocketed from 13% in 1960 to 35% in 2011-2012, with
most of this increase occurring since 1980 (Flegal et al., 1998; Ogden, et al. 2014). Obesity has
become a major public health and public finance concern. Estimates of its annual costs include
112,000 lives and $190 billion, with about half of the medical expenses borne by Medicare and
Medicaid (Flegal et al., 2005; Cawley and Meyerhoefer, 2012; Finkelstein et al., 2003).
This trend has prompted economists to ask whether obesity is an economic phenomenon
involving individuals’ responses to incentives. Technological progress has resulted in an
environment in which food is cheaper and more readily available, while physical activity is
increasingly easy to avoid. Philipson and Posner (1999) formalize this notion by modeling
weight as the result of eating and exercise decisions made through a utility-maximization
process.2 Individuals trade-off the disutility from excess weight with the enjoyment of eating and
having a sedentary lifestyle, subject to a budget constraint. The model predicts that lower food
prices and reduced on-the-job physical activity increase weight, while the effect of additional
income on weight varies across the income distribution. Cutler et al. (2003) point out that time
costs of eating should matter in addition to monetary costs, and discuss how innovations such as
vacuum packing, improved preservatives, and microwaves have reduced the time cost of food
preparation. Later theoretical models (e.g. Komlos, 2004; Ruhm, 2012; Courtemanche et al.,
2012) add an intertemporal dimension, noting that the enjoyment from eating and sedentary
activities occurs in the present but the health costs occur in the future. The prediction that the
1 BMI=weight in kilograms divided by height in squared meters.
2 The paper was later published as Philipson and Posner (2003), but we focus on the working paper version as it
contains a more detailed model.
2
weights of at least some individuals respond to economic incentives persists in these models,
regardless of whether or not preferences are time consistent.
Motivated by these theoretical considerations, a large number of empirical studies
investigate links between various economic factors and obesity.3 Lakdawalla and Philipson
(2002) document an inverted U-shaped association between income and BMI in individual fixed
effects models. Lindahl (2005) and Cawley et al. (2010) find no evidence that income affects
weight using lottery prizes and variations in Social Security payments as natural experiments,
while Schmeiser (2009) finds that Earned Income Tax Credit benefits increase weight.
Several papers document a connection between the costs of eating and BMI. Lakdawalla
and Philipson (2002), Chou et al. (2004), Lakdawalla et al. (2005), Goldman et al. (2011), and
Courtemanche et al. (2012) find an inverse association between food prices and obesity, while
the results from Baum and Chou (2011) and Finkelstein et al. (2012) are less clear. Evidence on
the role of restaurants is mixed. Chou et al. (2004), Rashad et al. (2006), Dunn (2008), and
Currie et al. (2010) find a positive relationship between restaurant prevalence and BMI; but
Anderson and Matsa (2011), Baum and Chou (2011), and Finkelstein et al. (2012) find no
evidence of a connection. Cutler et al. (2003) argues that lower time costs of food preparation are
partly responsible for trends in weight. Additionally, several studies investigate whether food
stamps lead to obesity, with mixed results.4
A variety of other economic factors have been linked to BMI. Chou et al. (2004), Baum
(2008) and Rashad et al. (2006) estimate that higher cigarette prices increase obesity; however,
Gruber and Frakes (2006) and Nonnemaker et al. (2008) find that this result disappears using
3 A separate but related literature studies how economic factors affect childhood obesity. Since our study focuses on
adult obesity, we do not discuss this literature. See Anderson and Butcher (2006) for a survey of this literature, and
Cawley and Ruhm (2011) for a detailed discussion of research on both adult and childhood obesity. 4 See Baum (2011), Baum and Chou (2011), Beydoun et al. (2008), Chen et al. (2005), Fan (2010), Gibson (2003
and 2006); Meyerhoefer and Pylypchuck (2008), Kaushal (2007), and Ver Ploeg et al. (2007).
3
different methodologies, and Courtemanche (2009b) and Wehby and Courtemanche (2011)
suggest the long-run relationship might even be negative. The effect of urban sprawl on obesity
is also the subject of debate, with Ewing et al. (2003), Frank et al. (2004), and Zhou and
Kaestner (2010) obtaining a positive relationship with obesity but Plantinga and Bernell (2007)
and Eid et al. (2008) arguing otherwise. Other factors that have been linked to adult obesity
include on-the-job physical activity (Lakdawalla and Philipson, 2002; Lakdawalla et al., 2005;
Baum and Chou, 2011), state unemployment rates (Ruhm, 2000 and 2005), work hours
(Courtemanche, 2009a), gasoline prices (Courtemanche, 2011), and the proliferation of Walmart
Supercenters (Courtemanche and Carden, 2011).
Most of the aforementioned papers examine only one or a few factors, and it is difficult
to use their results to answer the big-picture question of how well “the economic explanation” of
people responding to changing incentives can explain the rise in obesity. Simply adding the
percentage of the trend explained by separate studies of each potential contributor does not
produce a reliable answer. Many of the economic variables discussed above are highly correlated
with each other, so including only a small subset of them might lead to omitted variable bias.
Summing the effects of those variables would then lead to double counting some of their
contributions to the rise in obesity. For example, the number of stores selling food likely affects
food prices; so if one study estimates the impact of grocery stores while another estimates the
effect of food prices, the portion of food stores’ impact that occurs via prices will be double
counted. Other examples include the influences of restaurant density on restaurant prices, gas
prices on urban sprawl, and income on various aspects of the built environment. To underscore
our point, Table 1 shows that adding estimates from the literature suggests that economists have
already explained 177% of the rise in average BMI.
4
Chou et al. (2004) provide the first attempt at a comprehensive economic model of
obesity that includes several economic factors. They use the 1984-1999 Behavioral Risk Factor
Surveillance System (BRFSS) combined with state-level prices of grocery food, restaurant
meals, cigarettes, and alcohol as well as restaurant density and clean indoor air laws. In models
that control for individual demographic characteristics and state fixed effects, these state-level
economic factors explain essentially all of the growth in BMI and obesity during the period.
However, Chou et al. (2004) do not control for time in any way, which – as noted by Gruber and
Frakes (2006) and Nonnemaker et al. (2009) – likely introduces bias due to the strong upward
trend in weight. In the original working paper version of their work, Chou et al. (2002) show that
including a quadratic time trend leads to smaller coefficient estimates than those from models
without controls for time. When we estimate their model with our data (through 1999, the last
year of their sample), adding year fixed effects substantially attenuates the estimates. Appendix
Table 1 reports these results.
Recognizing this issue, two recent papers aim to develop comprehensive economic
models of obesity while controlling for time. Finkelstein et al. (2012) forecast obesity through
2030 based on a model that includes individual demographic characteristics as well as state-level
unemployment rate, alcohol price, gasoline price, fast food and grocery food prices, the relative
price of healthy to unhealthy foods, restaurant density, and internet access. They find scant
evidence that these state-level economic factors influence obesity. Baum and Chou (2011)
perform a Blinder-Oaxaca decomposition using data from the 1979 and 1997 cohorts of the
National Longitudinal Survey of Youth in an effort to explain the differences in BMI between
the two cohorts. They include economic factors related to employment, on-the-job physical
prevalence, but find that these variables explain very little of the rise in obesity, at least among
their sample of young adults.
We contribute to this literature by providing an analysis of body weight trends that is, to
our knowledge, the most comprehensive in terms of the number of economic factors included,
the length of the sample period, and the range of BMI-related outcomes considered. We combine
individual-level survey data from the 1990-2010 waves of the Behavioral Risk Factor
Surveillance System with 27 state-level variables reflecting general economic conditions; labor
supply; and the monetary or time costs of eating, physical activity, and smoking. Factors related
to general economic conditions include the unemployment rate, median income, and measures of
income inequality. Our labor supply variables are female and male labor force participation rates,
average work hours, and proportions of physically active and blue collar jobs. Factors
influencing the monetary or time costs of caloric intake include restaurant, grocery food, and
alcohol prices; the relative price of fruits and vegetables to other foods; restaurant,
supercenter/warehouse club, supermarket, convenience store, and general merchandiser
densities; and per-capita food stamp spending. Variables influencing the relative costs of
physical activity are gasoline prices, fitness center density, and a proxy for urban sprawl.
Cigarette prices and smoking bans capture variation in the costs of smoking.
We estimate how these economic factors are associated with BMI, obesity, and class
II/III obesity (BMI≥35, also known as severe obesity), as well as various percentiles of the BMI
distribution. Our models control for demographic characteristics as well as state and year fixed
effects. Changes in the economic factors collectively explain 37% of the rise in average BMI and
43%, 59% and 51% of the increases in obesity, class II/III (severe) obesity, and the 90th
percentile of the BMI distribution. The high explanatory power for the trends in severe obesity
6
and the 90th
BMI percentile is particularly important, as this is where the strong deleterious
mortality and morbidity consequences of excessive weight occur (Flegal et al., 2013).
Supercenter/warehouse club expansion and increasing numbers of restaurants are the leading
drivers of the results. The decline in blue collar employment and rise in food stamp spending
also explain meaningful portions of the trend in class II/III obesity, with other factors adding
small contributions for particular outcomes.
Robustness checks show that our conclusions remain similar if we drop insignificant
factors, use a quadratic trend instead of year fixed effects, allow for gradual effects, aggregate
the data, or use instrumental variables for the leading contributors to the trend. We conduct
falsification tests that suggest little connection between the key economic factors and other
health behaviors, consistent with a causal interpretation of our main results. We also find that
supercenter and warehouse club density is associated with a higher probability of weight loss
attempts. Since weight loss attempts can be considered an admission of past deviations from
utility-maximizing levels of weight (Ruhm, 2012), this suggests the effect of
supercenters/warehouse clubs on weight may be partly attributable to time inconsistency.
II. Analytical Framework and Econometric Model
We model weight (W) as a function of caloric intake (I), energy expenditure (E), and
metabolism (M):
Greater caloric intake increases weight, while greater energy expenditure and a faster
metabolism reduce weight. Smoking’s ( ) effects are multifaceted: nicotine stimulates the
metabolism and has appetite-suppressing properties that may reduce caloric intake, but smoking
diminishes lung capacity which may reduce physical activity (Courtemanche, 2009b). Caloric
7
intake, exercise, and smoking are in turn influenced by variables related to their monetary and
time costs ( ) as well as general economic ( ) and labor market (L) characteristics.
Therefore,
Substituting equations (2) through (5) into (1) yields
which simplifies to the reduced-form equation
Estimating the full structural model in (6) with a large number of aggregate-level
economic factors is not practical with available data. Datasets that contain sufficient sample sizes
to simultaneously analyze the effects of many state-level economic variables (like the BRFSS)
lack adequate information on the mechanisms (eating, exercise, and/or smoking) through which
these variables influence weight, while sources that contain sufficient information on the
mechanisms (e.g. the National Health and Nutrition Examination Surveys) are too small. Our
empirical analysis therefore focuses on the estimation of the reduced-form model given by (7).
Assuming a linear functional form for (7) yields the estimating equation
8
where i, j, and t index individuals, states, and years. W=BMI, a dummy for obesity (BMI≥30), a
dummy for class II/III (BMI≥35), or various percentiles of the BMI distribution.5 is a set of
controls that includes individual age and age squared; dummies for gender, race/ethnicity (black,
white, Hispanic, or other), marital status (single, married, divorced, or widowed), and education
(less than high school degree, high school degree, some college, or college degree); as well as
state population.6 and are state and year fixed effects.
consists of four variables reflective of general state economic characteristics:
unemployment rate, median income, and the ratios of the 90th
to the 50th
and the 50th
to 10th
percentiles of the earnings distribution.7 Theoretically, income could influence weight in either
direction. Expanding the budget set could raise food consumption and higher weight, or it could
reduce weight by causing substitution from cheap, energy-dense foods to more expensive,
healthy foods. Additional income could also reduce weight by increasing demand for health, as
higher wages increase the value of healthy time (Grossman, 1972). Lakdawalla and Philipson
(2002) documented an inverted U-shaped relationship between income and BMI, with additional
income increasing BMI at the low end of the distribution but decreasing it at the high end. The
non-linearity of this relationship suggests that central tendency might not be the only feature of
the income distribution that influences the weights of a state’s residents; variance (i.e. income
inequality) might also matter. We also include unemployment rates because higher state
5 We have verified that our conclusions are similar if we use logits or probits for the binary dependent variables
rather than linear probability models. We present linear probability model results as they are easier to interpret. 6 We control for population because some of our economic incentive variables are per capita, and we want to ensure
that any estimated effects of these variables can be attributed to the numerator rather than the denominator. 7 The BRFSS does contain a variable for respondents’ household income, but it only gives broad categories and is
top-coded at $75,000. Because of the top-coding, inflation-adjusting this variable suggests that average real income
dropped by over 20% during our sample period, which is inconsistent with other data sources and might therefore
misleadingly suggest that changes in real income have substantially contributed to the obesity trend. We therefore
control for income at the state level rather than the individual level. It is unlikely that this would bias our coefficient
estimates for the regressors of interest since they are also state level. Indeed, these estimates are very similar if we
use the BRFSS individual income measure rather than median state income.
9
unemployment has been linked to lower BMI, with the association not being explained by
income (Ruhm, 2005).
L consists of five state-level variables related to labor supply: female and male labor
force participation rates, average work hours among employees, proportion with a job that
requires at least moderate physical activity (defined as a metabolic equivalent (MET) score of 3
or higher), and proportion of the workforce in blue collar occupations (construction,
manufacturing, or extraction). The first three of these reflect the impact of market work on time
constraints, perhaps leading to less exercise or substitution from home-cooked meals to less
healthy prepared foods. This theory is particularly salient in light of the rise in female labor force
participation during the 20th
Century that was only partially offset by a decline in male labor
force participation (Anderson et al., 2003; Ruhm, 2008; Courtemanche, 2009a). The latter two
variables relate to the notion that the shift from a manufacturing-based economy to more
sedentary jobs may have reduced overall levels of physical activity, as one must now exercise
during leisure time (Philipson and Posner, 2003; Lakdawalla and Philipson, 2005). Proportion in
active jobs captures this hypothesis more directly, while the share in blue collar occupations may
also capture other aspects of such jobs – e.g., their relatively rigid structure may inhibit on-the-
job snacking or going out for lunch.
includes several variables related to the monetary or time costs of calories. These
variables test a leading theory for the rise in obesity: that food has become cheaper and more
readily available, increasing caloric intake and therefore weight. The first three variables in this
category are restaurant, grocery food/non-alcoholic drink, and alcohol prices. At first glance,
lower prices for foods or drinks should increase weight via the law of demand; however,
substitution between types of food and drink needs to also be considered. For example, if the
10
price of grocery food falls while the price of restaurant meals stays the same, individuals might
substitute away from restaurant meals toward home-cooked meals, which are presumably less
caloric. Similar logic applies if the prices of certain types of grocery foods fall further than
others. To that end, our fourth variable in this category is the relative price of fruits and
vegetables to other grocery foods. Fifth, we include per capita food stamp spending, which
effectively lowers the price of food for recipients out to a certain threshold.
Our variables related to the time cost of obtaining food are per capita numbers of
restaurants, supercenters/warehouse clubs, supermarkets, convenience stores, and general
merchandisers. Greater availability of these stores reduces travel time to obtain food, presumably
increasing weight; however, substitutability matters here as well. For example, the food sold in
conventional supermarkets may be on average less energy-dense than food sold at the other
places. A rise in supermarket density could, therefore, reduce weight by lowering the time costs
of buying healthy foods. Food store availability could also influence monetary prices, either
through competitive effects or, in the case of supercenters and warehouse clubs, by selling food
at discounted prices (Courtemanche and Carden, 2011).
includes three state-level variables: gasoline price, fitness centers per capita, and share
of residents living in the central cities of MSAs. Higher gasoline prices increase the cost of
driving relative to walking, bicycling, or taking public transportation, effectively reducing the
opportunity cost of physical activity (Courtemanche, 2011).8 An increase in fitness center density
lowers the time cost of exercising. Share of residents living in central cities proxies for urban
sprawl.9 More sprawl (fewer residents in central cities) typically reduces the amenities accessible
8 Courtemanche (2011) notes that higher gasoline prices could also reduce eating at restaurants.
9 We considered other proxies for urban sprawl, such as population-weighted population density, and share of the
population living in counties with various density cutoffs. The conclusions were similar.
11
through walking or mass transit, increasing the opportunity cost of caloric expenditure (Zhou and
Kaestner, 2010).
Finally, includes state-level cigarette price and dummies for smoking bans in private
workplaces, government workplaces, restaurants, and other locations. Cigarette prices capture
the monetary cost of smoking, while smoking bans affect the time cost since smokers have to go
outside to smoke more often (Chou et al., 2004).
III. Data
Our source of individual-level data is the BRFSS, a telephone survey of the health
conditions and risky behaviors of randomly-selected individuals conducted by state health
departments and the Centers for Disease Control. The BRFSS began in 1984, but did not include
all states until the 1990s. We use the years 1990-2010 to match the years in which all of our
state-level economic factors are available. As already discussed, the sharp rise in obesity began
around 1980, so our sample includes two-thirds of the period during which weights rapidly
increased. Following Gruber and Frakes (2006), we exclude individuals older than 64 out of
concerns that the true model of weight for the elderly is likely different than that for working-age
adults, and that mortality is more likely endogenous to weight for seniors, which has implications
for the composition of the sample.
The BRFSS includes self-reported height and weight. We apply the percentile-based
correction of Courtemanche et al. (2014) to adjust for systematic reporting error, and use the
“corrected” heights and weight to compute BMI and indicators for obesity and severe obesity.
Like the more familiar approach discussed by Cawley (2004), this method uses external
validation samples drawn from the NHANES to predict measured weight and height; however,
percentile ranks of the self-reported variables, instead of the self-reports themselves, are used to
12
predict the actual measures. The resulting predictions are robust to differences in misreporting
between surveys.10
Finally, the BRFSS contains the individual-level demographic variables discussed above,
as well as questions on health behaviors that provide dependent variables for our falsification
tests. These include seatbelt use and utilization of three types of preventive medical care: flu
vaccinations (shot or spray), mammograms, and prostate screenings.
Our price data come from the Council for Community and Economic Research’s (C2ER)
Cost of Living Index (formerly known as the ACCRA Cost of Living Index). The C2ER Cost of
Living Index computes prices for a wide range of grocery, energy, transportation, housing, health
care, and other items in approximately 300 local markets per quarter throughout the US. Most of
these local markets are single cities, but some are combinations of cities or entire counties.
Following Chou et al. (2004), we average over the prices of each item in the given category (e.g.
grocery foods) for each market, weighting by the C2ER shares of each item’s importance in the
basket of goods. We then define state prices as the population-weighted average of the prices in
the state’s C2ER markets. Finally, we convert prices to 2010 dollars using the Consumer Price
Index for all urban consumers from the Bureau of Labor Statistics.
We use data from the Quarterly Census of Employment and Wages (QCEW) for the
numbers of restaurants, supermarkets, convenience stores, and general merchandisers in each
state. The data are collected by the BLS with the cooperation of the state agencies that manage
the Unemployment Insurance system. In our industries, the QCEW captures the universe of
establishments. The only missing values are due to BLS disclosure rules that protect
10
Courtemanche et al. (2014) find that misreporting is more severe in the BRFSS than the NHANES, as one would
expect given the differences in interview context. For example, NHANES respondents are interviewed in person, but
BRFSS respondents are interviewed by phone. We also allow for the possibility that misreporting varies over time
by matching samples from each year of the BRFSS to samples from the closest years of the NHANES.
13
confidentiality in small cells. The number of restaurants includes both fast food and full service.
When we model these two categories separately, we cannot reject the hypothesis that the effects
of both types are the same.11
The QCEW information on supercenters and warehouse clubs is missing for many
observations, so we construct this variable by updating the primary data collected by
Courtemanche and Carden (2011). The key limitation is that this variable only captures Walmart
Supercenters, Sam’s Clubs, Costcos, and BJ’s Wholesale Clubs. It does not, for instance, include
K-Mart or Target Supercenters. However, Walmart is by far the dominant supercenter chain,
while Sam’s Club, Costco, and BJ’s Wholesale Club are the only three major warehouse chains
operating in the U.S. We considered modeling Walmart Supercenters and warehouse clubs
separately but were unable to reject the hypothesis that their effects are the same.
The other state-level variables come from various sources. Median income,
unemployment rate, female and male labor force participation, proportion of the workforce in a
physically active and blue collar job, average work hours, and 90/50 and 50/10 ratios come from
the Current Population Study (CPS), which is conducted by the U.S. Census Bureau for the
Bureau of Labor Statistics. The United States Department of Agriculture provides information on
Supplemental Nutrition Assistance Program (food stamp) benefits. Population and share of the
population living in MSA central cities are taken from the U.S. Census Bureau. Cigarette prices,
inclusive of state and federal excise taxes, come from The Tax Burden on Tobacco (Orzechowski
and Walker, 2010).12
Finally, we construct dummy variables reflecting the extent of state clean
indoor air laws using data from Impacteen and the classification scheme of the 1989 Surgeon
General’s Report (U.S. Department of Health and Human Services, 1989).
11
Chou et al. (2004) combined fast-food and full-service restaurants for the same reason. 12
The Tax Burden on Tobacco reports prices both including and excluding generic brands. Following Chou et al.
(2004), we use the series excluding generics to allow for greater comparability across the sample period.
14
We measure economic factors at the state rather than county level because the state is the
narrowest geographic level for which all determinants are available. The CPS variables are
available at the county level but can be unreliable because the samples are frequently quite small.
The C2ER price data have virtually no coverage of rural counties and only contain a subset of
urban counties. We are not aware of any county-level source of cigarette prices that is available
through our entire sample period, and the smoking ban variables reflect state laws. QCEW
establishment counts are often suppressed in small counties due to confidentiality concerns.13
Additionally, the BRFSS is only designed to be representative at the state level, and county
identifiers are not even available for all counties until the 1998 wave of the public-use data (or
1994 wave of the restricted data).
Combining all of these sources yields a final sample of 2,922,071 person-year
Smoking ban: other 0.057 (0.023)** 0.005 (0.014) Notes: Standard errors, heteroskedasticity-robust and clustered by state, are in parentheses. *** statistically
significant at 1% level; ** 5% level; * 10% level. All regressions include the control variables and state and year
fixed effects. BRFSS sampling weights are used. N=2,922,071.
45
Table 3 – Impacts of One Standard Deviation Increases in Economic Factors on P(Obese)
Separate Regressions All Factors Together
General Economic Indicators
Unemployment rate 0.002 (0.002) -0.001 (0.001)
Median household income -0.001 (0.002) 0.003 (0.002)
90/50 ratio -0.003 (0.001)*** -0.0004 (0.001)
50/10 ratio -0.004 (0.001)*** -0.002 (0.001)**
Labor Supply Variables
Female labor force participation rate -0.002 (0.002) -0.002 (0.001)
Male labor force participation rate -0.001 (0.002) 0.001 (0.002)
Average work hours 0.0004 (0.001) -0.001 (0.001)
Proportion active job -0.004 (0.002)* -0.002 (0.002)
Proportion blue collar -0.003 (0.002) -0.001 (0.001)
Variables Related to Monetary or Time Costs of Calorie Intake
Smoking ban: other 0.1% (0.4%) 0.8% (0.5%)* 0.7% (0.9%)
Subtotal 4.0% (3.4%) 4.3% (4.3%) -1.0% (4.2%)
Total from Economic Factors 37.2% (10.6%)*** 42.8% (12.9%)*** 59.3% (16.9%)***
Total from Controls 10.4% (1.1%)*** 6.1% (1.3%)*** 2.7% (1.8%) Notes: Standard errors, heteroskedasticity-robust and clustered by state, are in parentheses. *** indicates statistically
significant at the 1% level; ** 5% level; * 10% level. BRFSS sampling weights are used.
48
Table 6 – Impacts of One Standard Dev. Increases in Economic Factors on BMI Quantiles
Smoking ban: other -0.025 (0.013)* -0.021 (0.013)* 0.002 (0.014) 0.028 (0.021) 0.027 (0.037) Notes: Standard errors are in parentheses; the variance-covariance matrix of the unconditional quantile regression model is estimated using 500 bootstrap
replications. *** indicates statistically significant at the 1% level; ** 5% level; * 10% level. All regressions include the control variables and state and year fixed
effects. BRFSS sampling weights are used. N=2,922,071.
50
Table 7 – Percentage of Rise in BMI Quantiles Explained
0.1 0.25 0.5 0.75 0.9
General Economic Indicators
Unemployment rate 4.3% 5.4%** 1.4% 0.2% -1.8%
Median household income 0.4% 0.4% 0.7%*** 0.3% 0.3%
90/50 ratio -0.9% -0.9% -0.1% -0.5% -0.9%
50/10 ratio 2.4% 0.03% 1.4%* 1.6%** 0.4%
Subtotal 6.4% 5.0%* 3.4% 1.8% -2.1%
Labor Supply Variables
Female labor force participation rate -1.7%*** -1.1%*** -1.1%*** -0.2% 0.4%
Male labor force participation rate -2.4% -1.1% -1.8%** -0.6% 2.3%**
Average work hours 0.5% -1.4%** -0.4% 0.6% -0.1%
Proportion active job -1.4% -1.1% 0.2% 0.6% -1.1%
Proportion blue collar 4.6% 2.1% 3.8%*** 1.0% 5.3%***
Subtotal -1.3% -2.7% 0.7% 1.4% 6.8%***
Variables Related to Monetary or Time Costs of Calorie Intake
Variables Related to Monetary or Time Costs of Physical Activity
Gasoline price -0.014 (0.022) -0.020 (0.023) --
Fitness centers 0.011 (0.015) 0.004 (0.013) --
Proportion central city -0.005 (0.013) -- --
Variables Related to Monetary or Time Costs of Smoking
Cigarette price 0.007 (0.020) -- --
Smoking ban: private -0.001 (0.005) -- --
Smoking ban: government 0.0003 (0.004) -- --
Smoking ban: restaurant 0.001 (0.004) -- --
Smoking ban: other -0.006 (0.005) -- --
Sample Size 515,116 515,116 515,116 Notes: Standard errors, heteroskedasticity-robust and clustered by state, are in parentheses. *** indicates statistically
significant at the 1% level; ** 5% level; * 10% level. BRFSS sampling weights are used.
56
Appendix Table A1 – Replications of Chou et al.’s (2004) Model for BMI
Smoking ban: other 0.054 (0.056) 0.253 (0.060)*** 0.020 (0.037)
Observations 1,111,074 912,454 912,454 Notes: Standard errors, heteroskedasticity-robust and clustered at the state*year level, are in parentheses. ***
indicates statistically significant at the 1% level; ** 5% level; * 10% level. Regressions include state fixed effects
and individual-level control variables for age, age squared, real income, real income squared, and dummies for male,
race/ethnicity (black, white, Hispanic, or other), marital status (single, married, divorced, or widowed), and
education (less than high school degree, high school degree, some college, or college degree). Chou et al. also
included full-service restaurant price and its square, but the variable was only available every five years and was
imputed for the other years. Perhaps for this reason, its effect was one of the weakest Chou et al. estimated. We have
not been able to find an annual measure and therefore do not include full-service restaurant prices in our dataset.
57
Appendix Table A2 – Variable Descriptions and Summary Statistics
Variable Source Description Mean (Standard
Deviation)
1990
Mean
2010
Mean
BMI BRFSS Body mass index 26.618 (6.141) 26.027 28.507
Obese BRFSS Dummy for BMI≥30 0.279 (0.449) 0.184 0.339
Severely obese BRFSS Dummy for BMI≥35 0.111 (0.314) 0.066 0.141
Black BRFSS Dummy for race/ethnicity is non-Hispanic black 0.100 (0.300) 0.100 0.104
Hispanic BRFSS Dummy for race/ethnicity is Hispanic 0.120 (0.325) 0.083 0.142
Other BRFSS Dummy for race/ethnicity is not white, black, or Hispanic 0.056 (0.229) 0.030 0.076
Male BRFSS Dummy for sex is male 0.519 (0.500) 0.509 0.520
Some high school BRFSS Dummy for some high school but no degree 0.071 (0.257) 0.093 0.058
High school graduate BRFSS Dummy for high school degree but no college 0.301 (0.459) 0.347 0.260
Some college BRFSS Dummy for some college but no four-year degree 0.282 (0.450) 0.275 0.268
College graduate BRFSS Dummy for college graduate or further 0.315 (0.464) 0.250 0.387
Married BRFSS Dummy for married 0.611 (0.487) 0.618 0.639
Divorced BRFSS Dummy for divorced 0.122 (0.328) 0.111 0.112
Widowed BRFSS Dummy for widowed 0.019 (0.138) 0.022 0.017
Age BRFSS Age in years 39.634 (12.506) 37.623 41.983
Population Census State population (in 10,000s) 12.694 (10.117) 11.557 13.941
Proportion central city Census Proportion of residents in central city of an MSA 0.254 (0.106) 0.273 0.246
Cigarette price Tax Burden
on Tobacco
Weighted average price of pack of cigarettes (2010$) 4.159 (1.318) 2.756 6.265
Smoking ban: private ImpacTeen Dummy for state law prohibiting smoking in private
workplaces
0.143 (0.351) 0 0.471
Smoking ban:
government ImpacTeen Dummy for state law prohibiting smoking in
government workplaces
0.170 (0.376) 0.007 0.521
Smoking ban:
restaurant ImpacTeen Dummy for state law prohibiting smoking in
restaurants
0.243 (0.429) 0 0.621
Smoking ban: other ImpacTeen Dummy for other state smoking bans 0.717 (0.450) 0.547 0.851
Notes: n=2,922,071 in all years, 55,922 in 1990, and 239,215 in 2010. BRFSS sampling weights are used. a indicates variable not available in 1999-2001, 2003-
2005, 2007, and 2009; b not available 2003-2004,
c not available 1990-1992,
d not available 1990-2000,
e only available in 1994, 1996, 1998, 2000, and 2003. If
variables are not available in 1990 their values in the first year they are available are reported in the “1990 mean” column.
60
Appendix Table A3 – Percentage of Rises in BMI, Obesity, and Severe Obesity Explained
by Changes in Controls
BMI Obesity Class II/III Obesity
Age 61.2% (1.0%)*** 48.0% (1.3%)*** 54.6% (1.7%)***
Age squared -42.8% (0.9%)*** -32.1% (1.2%)*** -39.4% (1.5%)***
Subtotal from Age 18.4% (0.3%)*** 15.9% (0.3%)*** 15.3% (0.4%)***
Some high school -0.2% (0.1%) -0.2% (0.1%)** -0.4% (0.3%)
High school graduate 0.6% (0.3%) 0.8% (0.3%)*** 1.2% (0.9%)
Some college 0.1% (0.03%) 0.1% (0.03%)*** 0.2% (0.1%)**
College graduate -8.1% (0.4%)*** -9.2% (0.4%)*** -11.0% (1.3%)***
Subtotal from Education -7.7% (0.2%)*** -8.4% (0.2%)*** -10.0% (0.2%)***
Black 0.3% (0.01%)*** 0.3% (0.01)*** 0.4% (0.01%)***
State population -0.9% (1.0%) -1.3% (1.3%) -1.0% (1.7%)
Subtotal from Other Controls -0.3% (1.0%) -1.3% (1.3%) -2.6% (1.6%)
Total from Controls 10.4% (1.1%)*** 6.1% (1.3%)*** 2.6% (1.8%) Notes: Standard errors, heteroskedasticity-robust and clustered by state, are in parentheses. *** indicates statistically
significant at the 1% level; ** 5% level; * 10% level. BRFSS sampling weights are used. + indicates age is modeled
using a quadratic specification; the reported percentage of the trends explained is the sum of the percentages